Learning Temporal Information for States and Events

Abstract

Knowing the typical duration of events (for example, hurricanes last hour or days but not seconds or years) supports a variety of tasks in automated machine reading. Recently, methods to learn these durations for a limited class have been reported. However, events are associated with several other typical times, such as initiation points and preparation intervals. In this paper we define six temporally related aspects of events. We describe an automated method to learn events from the web and patterns that signal the typical temporal characteristics of the events. Finally, we show which patterns tend to signal which aspects. This diversity of event types, temporal aspects, and time characteristics has never yet been reported.